Question: how to distinguish between a tsumogiri action and a non-tsumogiri action

[Cryolite]

I have a question about the Action class in the Python API.

Its constructor has the following signature:

def __init__(self, action_type: ActionType, tile: int = 0, consume_tiles: list[int] = [])

However, from this signature alone, I cannot tell how to distinguish between a tsumogiri action and a non-tsumogiri (tedashi) action when constructing an Action object.

Could you clarify how these two cases are supposed to be represented?

[smly]

Thank you for the question!

The Action class itself does not have an explicit tsumogiri flag, but the environment internally determines tsumogiri vs. tedashi automatically based on the tile ID.

In RiichiEnv, the tile field in Action uses a 136-ID scheme where each of the 4 physical copies of a tile has a distinct ID (e.g., tiles 0–3 are the four copies of 1m). Because of this, when you pass an Action(ActionType.DISCARD, tile=X) to env.step(), the engine compares X against the current drawn_tile:

• action.tile == drawn_tile → tsumogiri
• action.tile != drawn_tile → tedashi

Limitation

Note that Action.encode() maps tile IDs to a 34-type space (tile / 4), collapsing all 4 copies into a single ID. This means the encoded action space cannot distinguish:

• Tsumogiri vs. tedashi
• Red dora vs. normal tiles

These distinctions are relevant for training ML models, and model developers need to handle them separately (e.g., by adding extra features to the observation or augmenting the action representation. Related: https://github.com/smly/RiichiEnv/blob/main/docs/ENCODING.md).

To clarify these points, I will also add docstrings to the Action class. 👍

[Apricot-S]

@smly
In the current RiichiEnv implementation, is it correct that kara-giri (空切り; discarding a tile identical to the drawn tile from the hand, rather than discarding the drawn tile itself) is not possible?

As a strategic option in mahjong, there are situations where intentionally performing kara-giri can be advantageous.

[smly]

@Apricot-S Yes, kara-giri is possible in the current implementation because of the 136-ID scheme. Since each physical copy of a tile has a unique ID, you can choose which specific copy to discard.

You can identify which tile IDs are in your hand via obs.hand (a list of 136-ID tile IDs), and which tile was just drawn via the draw event, allowing you to make this distinction at decision time.